[PATCH] D87419: AMDGPU: Skip all meta instructions in hazard recognizer

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: rampitec, kerbowa.
Herald added subscribers: hiraditya, t-tye, tpr, dstuttard, yaxunl, nhaehnle, jvesely, kzhuravl.
Herald added a project: LLVM.
arsenm requested review of this revision.
Herald added a subscriber: wdng.

This was not adding a necessary nop due to thinking the kill counted.


https://reviews.llvm.org/D87419

Files:
  llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp
  llvm/test/CodeGen/AMDGPU/hazard-recognizer-meta-insts.mir


Index: llvm/test/CodeGen/AMDGPU/hazard-recognizer-meta-insts.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/hazard-recognizer-meta-insts.mir
@@ -0,0 +1,41 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=gfx906 -run-pass=post-RA-hazard-rec -o - %s | FileCheck -check-prefix=GFX9 %s
+
+# Make sure the kill is skipped for hazard purposes, so the nop is
+# correctly inserted.
+
+---
+
+name: global_store_dwordx4_data_hazard_kill
+
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1, $vgpr2_vgpr3_vgpr4_vgpr5
+    ; GFX9-LABEL: name: global_store_dwordx4_data_hazard_kill
+    ; GFX9: GLOBAL_STORE_DWORDX4 $vgpr0_vgpr1, $vgpr2_vgpr3_vgpr4_vgpr5, 0, 0, 0, 0, implicit $exec
+    ; GFX9: $vgpr2 = KILL
+    ; GFX9: S_NOP 0
+    ; GFX9: $vgpr2 = V_MOV_B32_e32 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 $vgpr0_vgpr1, $vgpr2_vgpr3_vgpr4_vgpr5, 0, 0, 0, 0, implicit $exec
+    $vgpr2 = KILL
+    $vgpr2 = V_MOV_B32_e32 0, implicit $exec
+
+...
+
+---
+
+name: global_store_dwordx3_data_hazard_kill
+
+body:             |
+  bb.0:
+    liveins: $vgpr0_vgpr1, $vgpr2_vgpr3_vgpr4
+    ; GFX9-LABEL: name: global_store_dwordx3_data_hazard_kill
+    ; GFX9: GLOBAL_STORE_DWORDX3 $vgpr0_vgpr1, $vgpr2_vgpr3_vgpr4, 0, 0, 0, 0, implicit $exec
+    ; GFX9: $vgpr2 = KILL
+    ; GFX9: S_NOP 0
+    ; GFX9: $vgpr2 = V_MOV_B32_e32 0, implicit $exec
+    GLOBAL_STORE_DWORDX3 $vgpr0_vgpr1, $vgpr2_vgpr3_vgpr4, 0, 0, 0, 0, implicit $exec
+    $vgpr2 = KILL
+    $vgpr2 = V_MOV_B32_e32 0, implicit $exec
+
+...
Index: llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp
+++ llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp
@@ -368,7 +368,7 @@
     if (IsHazard(&*I))
       return WaitStates;
 
-    if (I->isInlineAsm() || I->isImplicitDef() || I->isDebugInstr())
+    if (I->isInlineAsm() || I->isMetaInstruction())
       continue;
 
     WaitStates += SIInstrInfo::getNumWaitStates(*I);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D87419.290834.patch
Type: text/x-patch
Size: 2122 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200909/d9f2c0cc/attachment.bin>